AI NewsInfrastructureAnnouncement
Hugging Face moves its model router to OpenRouter and its 80 providers
Hugging Face now sends every Inference Providers call through OpenRouter. One token reaches more than 80 providers.

Why it mattersA team calling an open-weight model on Hugging Face no longer picks a provider. The router retries on another if the first fails.
A team that calls an open-weight model through Hugging Face used to pick from a short list of providers. The list is now 80 plus. The choice is made for every call. Hugging Face says the choice uses the live price, speed and reliability of each one.
Hugging Face wrote on 16 November that it has switched Inference Providers "from our own custom router to OpenRouter's". Every request now goes through OpenRouter. If the first provider fails, the router tries another in the same call. The author is the Hugging Face Inference Providers team.
What a user gets
Hugging Face says every open-weight model on the Hub can be called with your Hugging Face token. The call goes to any provider that serves that model. The pool sits behind the OpenAI-shaped and the Anthropic-shaped API. Client code that already speaks either one needs no change. A user can also pin a model to one provider or one mode. The pin is a short suffix after the model id, for cheapest, fastest or a named provider.
The three numbers Hugging Face publishes
Hugging Face says OpenRouter's routing cut capacity errors by 46 percent. It says time to first token went up by 16.9 percent. It says throughput went up by 16.3 percent. Each number is measured on Hugging Face's own traffic. The post does not say the baseline, the window or the method. These are Hugging Face's figures for its own change.
Hugging Face also publishes three scale numbers. The combined service has 80 plus providers and 600 plus models. It routes more than 800 trillion tokens each month. The share of those tokens going to open-weight models was 26.8 percent in October 2025. It was 71.7 percent in September 2026. These are its own measurements.
Why Hugging Face moved
Hugging Face says building a router for 80 plus providers "would be a full-time job". OpenRouter already does it. Hugging Face keeps credits, catalogue and billing on its side. The price a user pays for a model does not change. OpenRouter's catalogue also carries closed-source models, which Hugging Face does not include in Inference Providers.
For a team using the service
A team that wants the cheapest or fastest provider on every call now gets it by writing one suffix. A team that wrote its own retry code can remove it. The router retries another provider when the first one fails. A team paying for a closed-source model through Hugging Face is unaffected. This change is for open-weight models only.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


